<p>Robust dynamic walking of humanoid robots is the foundation for their application in complex scenarios. However, the dynamic movement of humanoid robots on complex terrains (such as uneven ground and stairs) is confronted with challenges of balance maintenance, contact constraints and high computational costs. This paper presents a complete planning and control framework to achieve dynamic stable walking of humanoid robots on complex terrains. Based on biological research, the upper-level planner employs a unified MPC framework of 3D Divergent Component of Motion (DCM) for bio-inspired gait generation of humanoid robots, simulating human walking processes in complex terrains. The lower-level controller utilizes a Whole-Body Control (WBC) scheme based on task-space inverse dynamics to optimize the joint torques of the robot. The proposed framework enhances the dynamic performance and robustness of humanoid gait control through decoupled optimization and adaptive frequency adjustment, while maintaining simple yet efficient dimensional scalability. The proposed MPC framework decouples the optimization into three directions. During level-ground walking where the Center of Mass (CoM) height remains constant, the Quadratic Programming (QP) of z-direction is excluded from the optimization process. The approach simplifies the optimization process and reduces computational costs while ensuring real-time system performance. When the robot traverses uneven terrain or climbs stairs, QP-z dynamically optimizes the CoM height in real-time, adaptively adjusts the natural frequency of DCM, and outputs the results to the QP-x and QP-y for coordinated control. Simulation and experiments with humanoid robot ’Dexbot’ demonstrate that proposed control framework significantly enhances the adaptability of humanoid robots in complex terrains, reduces the cost of controller adjustments during terrain transitions, and enables stable dynamic walking across various terrains.</p>